[11a/n] Migrate Marlin kernels to torch stable ABI (#45176)

Signed-off-by: Chris Leonard <chleonar@redhat.com>
This commit is contained in:
Chris Leonard
2026-06-11 21:22:37 -07:00
committed by GitHub
parent 226ba9fc9e
commit 7021be66e8
19 changed files with 626 additions and 572 deletions
-29
View File
@@ -101,35 +101,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
") -> Tensor");
// conditionally compiled so impl registration is in source file
// Marlin Optimized Quantized GEMM (supports GPTQ, AWQ, FP8, NVFP4, MXFP4).
ops.def(
"marlin_gemm(Tensor a, Tensor? c_or_none, Tensor b_q_weight, "
"Tensor? b_bias_or_none,Tensor b_scales, "
"Tensor? a_scales, Tensor? global_scale, Tensor? b_zeros_or_none, "
"Tensor? "
"g_idx_or_none, Tensor? perm_or_none, Tensor workspace, int b_type_id, "
"SymInt size_m, SymInt size_n, SymInt size_k, bool is_k_full, "
"bool use_atomic_add, bool use_fp32_reduce, bool is_zp_float) -> Tensor");
// conditionally compiled so impl registration is in source file
// gptq_marlin repack from GPTQ.
ops.def(
"gptq_marlin_repack(Tensor b_q_weight, Tensor perm, "
"SymInt size_k, SymInt size_n, int num_bits, bool is_a_8bit) -> Tensor");
// conditionally compiled so impl registrations are in source file
// awq_marlin repack from AWQ.
ops.def(
"awq_marlin_repack(Tensor b_q_weight, SymInt size_k, "
"SymInt size_n, int num_bits, bool is_a_8bit) -> Tensor");
// conditionally compiled so impl registrations are in source file
// preprocess W-int4A-fp8 weight for marlin kernel
ops.def(
"marlin_int4_fp8_preprocess(Tensor qweight, "
"Tensor? qzeros_or_none, bool inplace) -> Tensor");
// conditionally compiled so impl registrations are in source file
#endif
}